Related Experiment Videos
CLEAR-AI: confounder-aware learning for equitable and accurate reasoning in AI for diagnosis
Vedant Joshi1, Ramon Correa1, Avisha Das2
1Arizona State University, School of Computing and Augmented Intelligence, Tempe, Arizona, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|May 25, 2026
Summary
We developed an AI debiasing method to reduce healthcare disparities across sensitive attributes like age and race. This approach ensures fairness in AI diagnostic models without compromising performance.
Area of Science:
- Artificial Intelligence in Healthcare
- Medical Imaging Analysis
- Algorithmic Fairness
Background:
- Deployment of AI in healthcare faces challenges due to implicit bias in models.
- Bias often targets multiple, correlated sensitive attributes, leading to health disparities.
Purpose of the Study:
- To develop and evaluate a novel AI approach for debiasing diagnostic models.
- To simultaneously address bias across multiple correlated sensitive attributes in healthcare AI.
Main Methods:
- A multibranch adversarial debiasing approach was developed.
- Utilized dynamic weighted gradient reversal for debiasing.
- Evaluated on chest X-ray data from CheXpert, MIMIC-CXR, and Emory Healthcare datasets.
Main Results:
- The debiased model showed comparable disease classification performance to baseline models.
- Reduced disparities in true positive rate (TPR) and false positive rate (FPR) across age, race, and support device subgroups.
- Maintained overall task performance after debiasing.
Conclusions:
- The proposed adversarial debiasing framework effectively reduces disparities in AI models.
- The methodology is extendable to numerous sensitive variables.
- Adversarial training shows potential for enhancing fairness and trustworthiness of AI in healthcare.
